How Big is the Media Multiplier? Evidence from Dyadic News Data Timothy Besley Thiemo Fetzer Hannes Mueller∗ August 27, 2020 Abstract This paper uses novel data to show how the media amplifies the economic im- pact of newsworthy events - the media multiplier. Specifically, we combine monthly aggregated and anonymized card spending data from 114 card issuing countries in 5 destination countries (Turkey, Egypt, Tunisia, Israel and Morocco) with a large corpus of news coverage of violent events in these destinations. To define and quantify the media multiplier we estimate a model of how media coverage helps shape beliefs about risks. When a country is perceived as dangerous by all potential visitors, aggregate spending falls by 53 percent with more than half of this effect due to the media multiplier. JEL Classification: O1, F5, F1, L8 Keywords: media, economic behavior, ∗Corresponding author: Mueller, [email protected]. Besley is based at the London School of Economics, Fetzer at Warwick University and Mueller is based at IAE(CSIC) and the Barcelona GSE. We thank audiences at Paris School of Economics, CAGE/Warwick University, Banco de Espana, CIREQ in Montreal, Barcelona GSE Summer Forum, London School of Economics, CUNEF Madrid. We are grateful to the International Growth Centre (IGC) and to the Mastercard Center for Inclusive Growth for their support. Mueller acknowledges fiancial support from grant PGC2018-096133-B-100 by MCIU/AEI/FEDER, grant 2017 SGR 1359 by AGAUR-GENERALITAT DE CATALUNYA and the Spanish Ministry of Economy and Competitiveness, through the Severo Ochoa Programme for Centres of Excellence in R&D (SEV-2015-0563) and Grant number ECO2015-66883-P. We thank Chris Dann, Olga Starchikova and Bruno Conte Leite for excellent research assistance. All errors are our own. 1 Introduction There is strong evidence that media reporting can have an impact on economic behavior.1 A case in point are negative events like crime, violence, traffic accidents or a pandemic where risk assessments and responses may be driven by media coverage.2 In such cases, the economic consequences of negative shocks could be magnified by media coverage something that we refer to as the media multiplier. Estimating the magnitude of the media multiplier requires a theoretically-grounded approach along with suitable data through which the economic impact of an event can be separated from the impact of media coverage of that event. This paper proposes an approach to estimating the media multiplier. Instead of relying exclusively on the timing of news (as for example, in Bloom, 2009 and Ramey, 2011) we use dyadic data, exploiting the fact that people travel to a destination country from different points of origin and are therefore exposed to different news coverage of the same events. Our approach relies on a novel data set that combines monthly aggregated and anonymized card spending data3 from 114 origin countries in 5 travel destinations (Turkey, Egypt, Tunisia, Israel and Morocco) with a large corpus of news coverage of the destinations at the different origins. Supervised machine learning allows us to capture the intensity of media coverage of violent events in this large news corpus and we are therefore able to study the response of travel to the media coverage of the same event across different origins. The kind of data available to us is illustrated in Figure1 which shows the pattern of card spending around a specific violent incident in which thirty-nine tourists were killed on June 26th 2015 in Sousse, Tunisia. The majority of the victims were UK citizens. The lines in the figure contrast the response of tourism activity on two specific dyads based on origin country: British (the dotted line) and German (the dashed line), while overall average spending patterns across all dyads is indicated by 1See Stromberg¨ , 2015 and Prat and Stromberg¨ (2011) for reviews. 2In 2016, terrorism in the US caused less than 0.01% of all deaths but was covered by newspapers more than any other cause of death – including the main causes: cancer and heart disease. See Combs and Slovic(1979) and Shen et al.(2018) and data compiled in Our World in Data. 3Mastercard made the anonymized and aggregated data available subject to robust privacy and data protection controls and in line with their principles guiding the ethical collection, management and use of data. 2 the solid line. (Figure1) The figure highlights that overall tourism spending across origins falls immediately in the months following the attack. However, spending of British travelers dropped much more compared to that of Germans. This motivates exploring how differential intensity of news reporting on the event in the two origin countries is responsible for the heterogenous response. And indeed, robust reduced-form results show that the media matters over and above the violent events themselves. We show that, even when we control for all variation specific to dyads and all time variation at each destination and origin, card activity on a dyad changes distinctly with differential media coverage. However, to interpret this as a media multiplier requires an economic model of how beliefs are formed. We posit a model in which potential visitors to a country rely on the media to form their views and some people use exclusively news outlets in their home country even if they lead to biases in beliefs due to sensationalism, heterogeneity in media freedom, the political agendas or quality of these news outlets. This is compounded of the use of so-called availability heuristics, whereby the ease with which information about a phenomena is recalled, affects an individual’s estimate of the likelihood of it recurring.4 Once we model the formation of beliefs, we can estimate the parameters of the updating process and show how such biased beliefs affect the size of the media mul- tiplier. Our model-based approach suggests that, when a country is perceived as dangerous by all potential visitors, card activity falls by 53 percent with more than half of this effect due to the media multiplier. Drilling down further, we show that visitors from countries with free media seem to rely much more on the news coverage in their country of origin. By building a theoretical model, we are able to provide a more convincing expla- nation of the heterogeneity and time path of media reporting than through a standard reduced-form approach. Moreover, it allows us to think about the implications of our 4See Tversky and Kahneman(1973). Recent research in psychology is reviewed in Zhu et al.(2020). 3 findings in wider contexts and to consider counterfactual paths. For example, we can use the distinction of coverage of violent events and background news which we gained from the supervised machine learning to show that news shocks are less se- vere in situations with more background news. The reason is that background news carry information and are therefore used in the updating process. The remainder of the paper is organized as follows. The next section reviews some relevant literature. In section3 we discuss the data used in some detail. Here, we also discuss the supervised learning method through which we make the text data usable for the subsequent analysis. In section4, we present reduced form results of the average effect of violence on tourism activity before incorporating the news data. In section5, we propose a statistical model and fit this on both the news and the tourism activity data. Concluding comments are in section6. 2 Related Literature This paper contributes to a large and growing literature which studies the impact of the media on economic and political outcomes.5 One of our contributions is to look at news events and coverage across multiple countries which complements a body of work that has focused in detail on US news coverage and its consequences. For example, Eisensee and Stromberg¨ (2007), shows that news on droughts affect US disaster relief. Similarly, Durante and Zhuravskaya(2018), suggests that offensives in the Israel Palestine conflict are strategically aligned to minimize news coverage in the US, while Jetter(2017) suggests that Al Qaida activity may be endogenous to preceding US television news coverage.6 We contribute to this literature by providing dyadic news data which we embed in a model of violence, reporting and beliefs. The paper provides a link between empirical work on the media and a large liter- ature in psychology which studies how individuals update their beliefs in response to news. Of particular relevance to our interpretation of the results is the large liter- 5The impact of the media on politics and economics has been the subject of many papers: Stromberg (2004) on redistributive spending, Besley and Burgess(2002) on government accountability, Gentzkow (2006); Bursztyn et al.(2017) on voter turnout, Snyder and Str omberg¨ (2010) on citizen knowledge, DellaVigna and Kaplan(2007); Enikolopov et al.(2011); Adena et al.(2015) on voting patterns, Durante et al.(2019) on the proclivity towards populist rhetoric. 6We observe that news coverage sharply responds to violent events, but not in anticipation. 4 ature in psychology focusing on availability heuristics, the human tendency to judge probabilities by the ease with which information underpinning them can be recalled Tversky and Kahneman(1973). 7 We offer a specific model which we fit to the data. In the model, individuals are assumed to be categorical thinkers and who update their beliefs in a partially na¨ıve fashion based on the news in their own country.8 Media coverage is an imperfect signal of the underlying categorical state and, as a result beliefs and economic behavior, “overreact” to it. Our idea of a media multiplier is closely related to an emerging literature on how news “shocks” affect economic behavior (see Ramey, 2011).